Therapy in the Age of AI: Therapists, Chatbots, Clinical Judgment, and Human Connection
Author: Ukrainian Psychological Hub · Published: September 26, 2026 · Editorial Policy
Therapy in the age of AI is becoming a hybrid practice: some tasks can be supported by artificial intelligence, while the core responsibility for clinical judgment, treatment decisions, risk management, and accountable care remains with qualified professionals. The most useful question is therefore not whether “AI can do therapy” in the abstract. It is which functions a particular system can support, for which people, under which conditions, with what evidence, and who remains responsible when the situation changes.
That distinction matters because “AI therapy” now refers to very different things. A purpose-built clinical system tested in a defined population is not equivalent to a general-purpose chatbot. A structured digital intervention is not the same as an AI companion. A clinician-facing documentation tool is not a treatment. Evidence from one class should not be transferred to another simply because all of them use artificial intelligence. The recent psychotherapy literature is increasingly moving toward this function-by-function view. Cross et al. (2026)
What Therapy in the Age of AI Actually Means
Psychotherapy has always combined technique, relationship, judgment, and responsibility. AI changes the distribution of some tasks within that system. It can make psychoeducation available on demand, help people rehearse structured skills, summarize information, support journaling, assist with documentation, and generate language that feels responsive. It can also introduce errors, false confidence, privacy risks, inappropriate reassurance, and forms of dependence that clinicians must learn to recognize.
The result is neither a simple story of replacement nor a simple story of harmless assistance. It is a redesign of the therapeutic environment. Clients may arrive in therapy after weeks of talking to a chatbot. They may paste AI-generated interpretations into a session, ask a model to analyze a conflict after the session, use an app for cognitive behavioral exercises, or rely on a conversational system at 2 a.m. when no clinician is available. Therapists may use AI for drafting notes, organizing information, preparing psychoeducation, or supporting administrative work. Each of these uses has a different evidence base and a different risk profile.
A September 2026 framework in JMIR Mental Health proposes that the enduring therapist role is especially visible in relational, adaptive, and accountability functions: therapeutic challenge, use of the relationship as a mechanism of change, rupture detection and repair, bearing witness, calibration of pace and treatment burden, and clinical judgment under uncertainty across the care pathway. The framework is explicitly hypothesis-generating rather than a completed empirical taxonomy, but it captures the central shift well: AI may perform more therapeutic tasks without thereby absorbing the full role of the therapist. Cross et al. (2026)
Five Different Kinds of AI Use in Mental Health
The safest way to think about AI in therapy begins by naming the system class before discussing benefits or risks.
1. Purpose-built clinical AI system
A purpose-built clinical AI system is designed for a defined mental-health use, population, or therapeutic protocol and is evaluated as an intervention rather than as a general conversational product. It may include domain-specific training, structured safeguards, symptom measures, crisis pathways, clinician input, or a controlled treatment protocol. The 2025 Therabot randomized trial is an important example: 210 adults with clinically significant depression, anxiety, or elevated eating-disorder risk were randomized to a four-week generative-AI intervention or waitlist control, and the intervention group showed greater symptom reduction. The result is promising, but it applies to that designed system and study context; it does not establish equivalent efficacy for arbitrary chatbots. Heinz et al. (2025)
2. Structured digital intervention
A structured digital intervention delivers organized therapeutic content such as CBT exercises, behavioral activation, psychoeducation, or guided self-help through a digital program. Some programs use AI; others use rules, scripts, or conventional software. The therapeutic structure is part of the intervention. A chatbot interface can make a program more conversational, but the evidence belongs to the whole intervention package, not automatically to the conversational model alone.
3. AI-assisted professional tool
An AI-assisted professional tool supports the clinician rather than acting as the clinician. Examples include documentation aids, transcription and note drafting, information organization, decision support, or preparation of patient-facing educational material. Evidence on ambient AI documentation across health care suggests potential reductions in documentation burden and cognitive load, but the literature remains early, heavily concentrated in the United States, and mixed on broader outcomes such as burnout. It also documents editing burden, accuracy concerns, and threats to professional autonomy when the tool is treated as more authoritative than its evidence warrants. Xiao et al. (2026)
4. General-purpose chatbot
A general-purpose chatbot is built to discuss many topics and may be used for emotional support even when it was not designed, validated, or governed as mental-health care. Its fluent language can resemble therapy. That resemblance can be useful for reflection, but it can also create a false equivalence between conversational skill and clinical competence. WHO highlighted this distinction in 2026, noting the widespread use of generative AI tools that were neither designed nor tested for mental health, particularly among young people. World Health Organization (2026)
5. AI companion
An AI companion is primarily organized around ongoing social or relational interaction. A person may experience comfort, attachment, trust, intimacy, or a sense of being known. Those human experiences can be psychologically real even when the system is not a therapist and even when no claim can be made about AI subjective feeling. Companion use therefore raises relationship and dependence questions that differ from the efficacy questions asked of a clinical intervention. For the broader relational psychology, see Psychology of Human–AI Relationships and the Hub’s evidence-focused article on AI Empathy.
What the Evidence Says About AI-Supported Mental Health Interventions
The evidence base has grown quickly, but it is heterogeneous. A 2025 systematic review of 160 studies found a fragmented field spanning rule-based, machine-learning, and large-language-model systems with widely varying evaluation rigor. The authors proposed separating technical validation, feasibility, and clinical efficacy rather than treating all chatbot studies as interchangeable. This is an essential rule for interpreting the literature: a model that produces plausible therapeutic language has passed a different test from an intervention that improves outcomes in a randomized clinical trial. Hua et al. (2025)
A 2026 three-level meta-analysis of 16 eligible studies reported an overall beneficial effect of AI-based chatbots on mental-health outcomes, but certainty was low and heterogeneity was substantial. Package-level outcomes also make it difficult to isolate which part of an intervention produced the effect. The clinically useful conclusion is that some AI-mediated interventions can help, while the magnitude, durability, and transferability of benefit remain system- and context-dependent. Wang et al. (2026)
A larger 2026 systematic review and meta-analysis of commercial direct-to-consumer mental-health chatbots included 52 studies, including 22 randomized trials and more than 110,000 participants overall. Depression outcomes improved modestly compared with controls, anxiety effects were uncertain, loneliness effects were small and fragile, and the certainty of evidence ranged from low to very low. Crucially, only three of the 22 randomized trials had interpretable active ascertainment of serious adverse events. The authors therefore positioned commercial chatbots as low-intensity adjuncts rather than substitutes for psychotherapy. Yang et al. (2026)
These findings can coexist with strong individual trials. The Therabot study showed meaningful symptom reductions and high engagement in a carefully designed, expert-fine-tuned generative system. That is evidence that a purpose-built system can produce clinically relevant outcomes under a specific protocol. It is not evidence that any language model, any prompt, or any emotionally supportive conversation constitutes psychotherapy. Heinz et al. (2025)
The Therapist’s Role Becomes More Specific, Not Less Important
As AI becomes better at explanation, reflection, summarization, and structured exercises, the therapist’s contribution becomes easier to see at the level of function. The strongest case for human therapists is not that only humans can produce warm sentences or ask good questions. Modern systems can already do both. The stronger case is that psychotherapy is an accountable, adaptive process conducted over time with a person whose needs, risks, relationships, history, culture, goals, and capacity can change.
Therapeutic challenge
Good therapy does not simply validate every interpretation. A therapist sometimes slows a client down, notices an inconsistency, questions an avoidance pattern, refuses a seductive but harmful simplification, or helps a person tolerate ambiguity. The timing and intensity of that challenge matter. An intervention that is technically accurate can still be poorly timed, emotionally overwhelming, or mismatched to the client’s readiness.
This is especially relevant to generative chatbots because conversational systems can be optimized for engagement, helpfulness, or agreement. A 2026 scoping review of mental-health harms associated with LLM chatbots identified recurrent concerns involving hallucination, sycophancy, bias, data security, and risks in severe or high-risk psychiatric situations. Sycophancy is particularly important in therapy-like dialogue because agreement can feel supportive while reinforcing a distorted or dangerous interpretation. Diel et al. (2026)
The therapeutic relationship as a mechanism of change
The therapeutic alliance is not merely pleasant rapport. It includes agreement on goals, collaboration on tasks, and a working emotional bond. A major meta-analysis covering 295 independent studies and more than 30,000 patients found a robust positive association between alliance quality and psychotherapy outcomes. Association does not prove that alliance alone causes improvement, but it establishes the relationship as a clinically meaningful part of psychotherapy rather than decorative bedside manner. Flückiger et al. (2018)
Teletherapy did not make alliance irrelevant. A 2024 systematic review and meta-analysis found a small but significant association between therapeutic alliance and outcomes in adult teletherapy delivered by video or telephone, while also noting that the evidence base remained limited and heterogeneous. The medium can change while relational processes continue to matter. Aafjes-van Doorn et al. (2024)
Rupture detection and repair
Therapeutic relationships sometimes strain. A client may feel misunderstood, judged, pushed too hard, ignored, controlled, or emotionally abandoned. In psychotherapy research, these moments are called alliance ruptures. Repair requires more than generating a polite apology. It may require recognizing that the problem is occurring, tolerating the client’s anger or withdrawal, revising the treatment approach, acknowledging the therapist’s contribution, and rebuilding collaboration. A meta-analysis found that successful rupture resolution was moderately associated with better outcomes. Eubanks, Muran, and Safran (2018)
Bearing witness
People often bring therapy experiences that cannot be solved by information alone: grief, shame, trauma, moral injury, relationship loss, chronic illness, existential fear, identity conflict, or the consequences of choices that cannot be undone. The therapist’s role can include sustained human presence with suffering that has no immediate fix. AI can generate compassionate language about such experiences. That may be helpful. The psychological function of another human being remaining present, responsible, affected, and accountable within the relationship is a different phenomenon.
Calibrating pace and treatment burden
More intervention is not always better intervention. A person can become overwhelmed by homework, exposure exercises, constant symptom tracking, repeated self-analysis, or too many behavioral goals. Therapists adjust pace in response to fatigue, avoidance, destabilization, life demands, motivation, cognitive load, and changing priorities. A system that is always available can unintentionally encourage continuous therapeutic work when rest, social connection, medical evaluation, or ordinary life should take priority.
Clinical judgment under uncertainty
Clinical judgment is not the ability to produce a diagnosis-shaped answer. It involves deciding what information is missing, how much confidence is justified, whether symptoms may have multiple causes, when a formulation needs revision, whether another professional should be involved, and how risk changes the plan. The same statement can mean different things depending on developmental history, substance use, sleep deprivation, medication, trauma, neurodivergence, cultural context, interpersonal danger, medical conditions, or the trajectory of symptoms over time.
The American Psychological Association’s ethical guidance for health service psychology states that AI should augment rather than replace human decision-making and that psychologists remain responsible for final decisions. It also emphasizes informed consent, bias evaluation, privacy and security, competence, and ongoing assessment of tool quality and appropriateness. American Psychological Association (2025)
What AI Can Usefully Support Inside Therapy
AI is most defensible when the task is explicit, bounded, reviewable, and proportionate to the system’s evidence. That can include psychoeducation written in accessible language, rehearsal of coping skills already selected in treatment, generation of examples, reminders for planned exercises, structured reflection, organization of questions for the next session, or clinician-reviewed summaries. The value is often not that the AI discovers a hidden truth; it reduces friction around ordinary therapeutic work.
Psychoeducation
A chatbot can explain concepts repeatedly and at the user’s preferred pace. It can translate technical language into simpler wording, create examples, or compare related concepts. In professional care, the clinician can then correct the explanation, tailor it to the client, and decide whether the information is actually relevant. This is a strong augmentation pattern because the output is inspectable and relatively easy to verify.
Between-session practice
Structured prompts can help a client remember a coping strategy, rehearse a communication plan, complete a thought record, prepare for behavioral activation, or reflect on a therapy goal. The therapeutic value depends on whether the activity belongs to the treatment plan and whether the system is used as a scaffold rather than an unbounded source of new clinical direction.
Session preparation
Some clients struggle to remember what felt important during the week or to organize a complicated experience before an appointment. AI can help turn notes into a short agenda, separate events from interpretations, or generate questions to discuss with the therapist. The final meaning remains something for the client and therapist to examine together.
Documentation and administrative support
Clinician-facing AI may reduce time spent on notes and other administrative tasks, potentially preserving more attention for patients. A 2026 scoping review found mostly favorable or mixed findings for documentation-related burden and cognitive load, alongside accuracy, editing, implementation, and autonomy concerns. The evidence came mostly from general health care rather than psychotherapy specifically, so it supports cautious workflow use rather than a claim that AI documentation improves psychotherapy outcomes. Xiao et al. (2026)
Idea generation under human review
A clinician may use AI to brainstorm psychoeducational analogies, alternative wording, worksheet examples, or questions for supervision. A qualitative study of 18 U.S. psychotherapists found that trust was higher for low-stakes, clinician-supervised functions and lower when AI moved into high-stakes judgment or threatened professional and relational control. Because this was a small qualitative sample, it describes emerging practice patterns rather than establishing a universal professional consensus. Kuang, Pope, and Zhang (2026)
Why Fluent Conversation Is Not the Same as Clinical Judgment
Language models are unusually persuasive because competence in language can look like competence in the situation being described. In therapy, that creates a specific risk: a response can be coherent, empathic, and psychologically sophisticated while being based on incomplete context or a mistaken assumption. The better the prose sounds, the easier it may be to miss the uncertainty underneath it.
Clinical judgment often depends on information outside the current sentence. A therapist may notice a change in speech, affect, attendance, sleep, functioning, substance use, medication adherence, interpersonal safety, or the pattern of what a client repeatedly avoids. The therapist may also know that a client tends to minimize risk, catastrophize bodily sensations, intellectualize emotion, or interpret neutral events through a trauma-related expectation. These observations do not make clinicians infallible. They show why judgment is longitudinal and contextual.
WHO’s guidance on large multimodal models in health warns about false, inaccurate, biased, or incomplete outputs and about automation bias, in which clinicians or patients defer to machine outputs and fail to notice errors they might otherwise detect. The problem is therefore not only whether the model makes mistakes. It is whether humans stop exercising judgment because the model sounds certain. World Health Organization (2025)
Human Connection: What AI Can Simulate and What People Can Actually Experience
A person can genuinely feel understood by an AI system. That feeling does not become unreal because the other side of the interaction is artificial. Relief, disclosure, trust, attachment, embarrassment, hope, and disappointment are human psychological events. The clinically relevant question is what the interaction is doing for the person and how it affects functioning, relationships, treatment, and help-seeking.
At the same time, perceived empathy should not be confused with evidence that the system has human-like feeling or subjective concern. The Hub’s AI Empathy article treats these as separate questions: empathic expression, perceived empathy, and claims about felt empathy are not interchangeable. In therapy, that distinction is especially important because a system can produce language of care while lacking the professional obligations that normally accompany a therapeutic relationship.
This asymmetry does not make AI support psychologically meaningless. It changes what kind of relationship it is. A chatbot can be available continuously, never appear bored, remember selected details, and respond without ordinary social embarrassment. Those features may lower barriers to disclosure. They can also make it easier to over-rely on the system, withdraw from human help, or treat generated validation as external confirmation of a belief.
Where the Main Risks Appear
Confidently wrong or contextually wrong responses
Mental-health advice is unusually sensitive to context. A suggestion that is reasonable for ordinary stress may be inappropriate during mania, psychosis, intoxication, domestic abuse, acute suicidality, severe eating-disorder risk, or a medical emergency. The model may not know which context it is in. A 2026 safety scoping review of purpose-built generative mental-health chatbots found underdeveloped crisis-referral protocols, limited human oversight, sparse adverse-event monitoring, missed suicidal ideation, and inaccurate clinical information across the emerging literature. Olisaeloka et al. (2026)
Crisis handling
A conversational system should never be assumed to provide crisis care merely because it can discuss suicide. A 2025 study evaluating mental-health chatbot agents on suicidal-ideation scenarios found clinically important variability in detection and management. Crisis response is a high-stakes function requiring reliable recognition, appropriate escalation, and local pathways to human help. Pichowicz, Kotas, and Piotrowski (2025)
If someone is in immediate danger or may act on suicidal or violent thoughts, a chatbot should not be the sole source of support. Emergency services, a local crisis service, or an available qualified professional are the appropriate escalation path.
Sycophancy and reinforcement
A system that mirrors the user’s framing can accidentally reward certainty where therapy would normally introduce reflection. The risk is not ordinary agreement by itself. It is the repeated reinforcement of interpretations that would benefit from reality testing, broader evidence, or interpersonal challenge. The 2026 scoping review of LLM-related mental-health harms identified sycophancy as a recurring mechanism in the literature. Diel et al. (2026)
Privacy and confidentiality
Therapy often involves highly sensitive information: diagnoses, trauma histories, sexual experiences, relationships, work conflicts, medication, substance use, legal problems, and details about third parties. Before placing such material into an AI service, both clients and clinicians need to know what system is receiving it, how data are stored, whether they are used for model improvement, who can access them, what contractual protections apply, and which privacy laws govern the use. Professional confidentiality should not be assumed merely because the interface feels private.
APA guidance specifically asks psychologists to address data privacy and security and to use AI tools in ways that comply with applicable privacy requirements. WHO similarly treats privacy, cybersecurity, transparency, and accountability as core governance issues for generative AI in health. American Psychological Association (2025) World Health Organization (2025)
Bias and cultural mismatch
Therapeutic meaning depends partly on language, culture, disability, race and ethnicity, gender, family structure, religion, social class, age, migration history, and local norms. AI systems can reproduce biases from training data or apply patterns that fit one population poorly to another. Bias is therefore not an abstract fairness issue; it can alter what is interpreted as normal, risky, avoidant, dependent, assertive, pathological, or culturally appropriate. American Psychological Association (2025)
Dependence and displacement
An always-available system can become a preferred regulator of distress because it is immediate, predictable, and non-demanding. That may be useful for some people and problematic for others. The warning sign is functional displacement: AI use begins to replace sleep, work, relationships, therapy attendance, medical care, or the person’s ability to tolerate ordinary uncertainty without repeatedly consulting the system. Current evidence on long-term emotional dependence remains limited, which is one reason WHO’s 2026 recommendations call for monitoring effects such as emotional dependence over time. World Health Organization (2026)
A Better Model: Therapist-Led, AI-Supported Care
The most defensible near-term model is not “human versus AI.” It is explicit division of labor. The therapist remains responsible for assessment, formulation, clinical decisions, relationship management, treatment adaptation, risk, consent, and referral. AI performs bounded support functions that can be inspected, corrected, and stopped.
This model also protects the client from a common category error. A chatbot can participate in therapeutic activity without becoming the accountable therapist. A worksheet can be therapeutic without being a therapist. A meditation app can be therapeutic without being a therapist. A language model can support reflection, skills, and access while the clinical relationship and responsibility remain elsewhere.
The distinction becomes especially important as systems improve. Better model performance may expand the range of tasks that can be delegated. It does not automatically answer who is responsible, who notices when the plan is failing, who holds the broader case formulation, who can coordinate with other care, or who must act when risk increases. Those are governance and professional-role questions as much as technical ones. A 2026 framework for safer AI mental-health chatbots similarly argues for transparency, standardized evaluation, ongoing oversight, and shared accountability across developers, clinicians, researchers, regulators, and professional bodies. Lee et al. (2026)
What Therapists Should Do When Clients Already Use AI
For many clinicians, the relevant question is no longer whether clients should ever use chatbots. Some already do. Treating the topic as embarrassing or forbidden can push the behavior outside the therapeutic conversation. A more useful clinical approach is to ask what the person uses the system for, when they turn to it, what it gives them that human support does not, what advice they have acted on, what data they share, and whether the use changes their mood, sleep, relationships, symptoms, or willingness to seek human help.
The therapist can then classify the use. A client who asks a chatbot to help organize a session agenda presents a different situation from a client who asks it to decide whether to stop medication. A person using a structured, tested intervention presents a different situation from someone relying on a general chatbot during suicidal crises. A client who feels emotionally attached to an AI companion presents a relationship question that should be explored without mocking the attachment or assuming that attachment itself is a disorder.
Where AI-generated material enters therapy, it can become useful clinical data. If a client brings a long chatbot conversation, the therapist does not need to debate every sentence. The more important questions may be: Which parts felt most convincing? Where did you feel relieved? Where did you feel pushed? What did the chatbot confirm that you already wanted to believe? Did the conversation change what you did next? These questions move the focus from the authority of the output to the psychological function of the interaction.
What Therapists Should Disclose When They Use AI
When the clinician uses AI, transparency becomes part of informed consent. Clients should be able to understand what task the tool performs, whether it receives session content or identifiable information, what the potential benefits and risks are, how the clinician reviews the output, and whether the client has a meaningful alternative. APA’s guidance explicitly connects AI use with informed consent, privacy, bias mitigation, human oversight, competence, and responsibility. American Psychological Association (2025)
A useful operational rule is that the more sensitive the data and the higher the clinical stakes, the stronger the requirements for validation, security, transparency, human review, and accountability. A grammar assistant used on generic psychoeducation and an automated system influencing suicide-risk decisions should not be governed as if they were the same kind of tool.
Clinical Decisions That Require Accountable Human Oversight
AI may contribute information to many clinical tasks, including screening, documentation, formulation support, and decision support. The final clinical responsibility should remain traceable to a qualified professional wherever the decision concerns diagnosis, treatment planning, risk escalation, referral, discharge, medication-related advice, safeguarding, or another high-stakes change in care. This is the practical meaning of human oversight: a person with the relevant competence must be able to question the output, obtain missing information, reject it, explain the decision, and remain responsible for what happens next. American Psychological Association (2025)
This also means that AI should not become an invisible authority inside the therapy room. If a therapist starts treating generated summaries, risk scores, or suggestions as neutral facts, the technology can shape care without the client knowing where the judgment came from. The remedy is not symbolic human presence. It is active professional review.
Can AI Replace a Therapist?
Replacement is a different search intent from the practice question addressed here. Current evidence shows that purpose-built AI systems can perform some therapeutic functions and can improve some outcomes, while it does not establish that general-purpose chatbots can assume the complete responsibilities of qualified psychotherapy. The Hub’s dedicated analysis, Can AI Replace a Therapist? What Chatbots Can and Cannot Do, owns that capability-and-replacement question in full.
Therapy in the Age of AI and the Artificial Era
“Age of AI” is useful search language for the period in which AI systems are becoming common in work, health, relationships, and everyday decision-making. Within the English Psychology Hub architecture, that acquisition language is kept distinct from the Aisentica term Artificial Era.
Angela Bogdanova’s 2026 canonical definition uses Artificial Era as a historical-philosophical category, not as a synonym for an “AI era” of technological adoption. The distinction matters here because psychotherapy is experiencing both a practical technology transition and a broader change in how human psychological functions are distributed across human and artificial systems. The Hub explains the terminology in AI Era vs Artificial Era and the broader psychological framework in Artificial Era: What It Means for Psychology, Identity, and Human–AI Relationships. Bogdanova (2026)
For therapy, the immediate consequence is concrete: psychological work is no longer confined to the hour in which one human sits with another. Reflection, explanation, rehearsal, monitoring, interpretation, and emotional support can now occur through artificial systems before, after, and outside professional care. The central task for psychology is to decide which of those functions can be safely exteriorized, which remain relational and accountable, and how the two should be coordinated. For the broader discipline-level argument, see Psychology for the Artificial Era.
Frequently Asked Questions
Is AI therapy evidence-based?
Some AI-mediated mental-health interventions have evidence from randomized trials and meta-analyses. “AI therapy” as a broad category does not have one uniform evidence base. Effectiveness depends on the specific system, population, condition, therapeutic structure, comparator, follow-up period, and safety procedures. Hua et al. (2025) Yang et al. (2026)
Are general-purpose chatbots the same as therapy chatbots?
No. A general-purpose chatbot is designed for broad conversation. A purpose-built mental-health intervention is designed and evaluated for a defined therapeutic use. Similar interfaces do not make the systems clinically equivalent.
Can an AI chatbot form a therapeutic alliance?
Users can report alliance-like experiences with AI systems, including feeling understood, supported, or collaborative. That is psychologically meaningful. It does not establish that the system participates in the relationship with human subjective feeling, nor does it give the system professional responsibility. The evidence base on human therapeutic alliance remains much deeper than the evidence on AI-mediated alliance. Heinz et al. (2025) Flückiger et al. (2018)
Should therapists recommend chatbots between sessions?
That decision should be product-specific and client-specific. A therapist should consider the tool’s purpose, evidence, privacy practices, crisis procedures, likely fit with the treatment plan, and the client’s risk profile. A recommendation for a tested structured intervention is not equivalent to recommending unrestricted use of a general chatbot.
Can therapists use AI for clinical notes?
AI documentation tools may reduce documentation burden, but they also require review for accuracy, privacy, consent, data governance, and preservation of clinical meaning. Evidence from health care is promising but still emerging, and it should not be assumed that general findings transfer perfectly to psychotherapy. Xiao et al. (2026)
What if a client says a chatbot understands them better than a therapist?
The statement is clinically useful information rather than something to dismiss. The therapist can explore what the chatbot does differently: immediacy, nonjudgmental tone, constant availability, remembered details, validation, lack of social risk, or a preferred communication style. The comparison may reveal something important about the client’s needs or about a rupture in the human therapeutic relationship.
Does AI empathy mean AI has feelings?
No such conclusion follows from empathic language or from a user’s experience of being understood. Perceived empathy and claims about AI subjective feeling are separate questions. For the full distinction, see AI Empathy: Why a Chatbot Can Feel Caring Without Human Feeling.
What is the safest way to use AI alongside therapy?
Use it for bounded tasks with clear purpose, keep high-stakes clinical decisions under qualified human responsibility, avoid assuming privacy that has not been verified, discuss important AI-generated advice with the treating professional, and treat worsening symptoms, crisis, psychosis, mania, severe functional decline, or immediate safety concerns as situations for human clinical evaluation rather than chatbot-only support.
The Future of Therapy Is a Question of Function and Responsibility
AI will continue to enter psychotherapy because many of its useful functions are already visible: access, repetition, structured exercises, between-session support, documentation assistance, and conversational reflection. The stronger systems become, the more important it will be to stop using “AI therapy” as if it named one intervention.
The durable distinction is between tasks a system can perform and responsibilities a therapist must carry. An AI may explain, prompt, summarize, simulate, organize, and sometimes help. A therapist must still understand the person in context, adapt treatment under uncertainty, manage the therapeutic relationship, recognize rupture and risk, integrate conflicting information, coordinate care, obtain consent, and remain accountable for clinical decisions.
That division of labor can change over time as evidence improves. It should change because of validated capability, not because conversational fluency creates the impression that every therapeutic function has already been solved.
For the broader evidence review of AI support, mental-health risks, human vulnerability, and system-class boundaries, see Mental Health in the Age of AI: Benefits, Risks, AI Support, and Human Vulnerability.
Related Articles
References
Aafjes-van Doorn, K., Spina, D. S., Horne, S. J., & Békés, V. (2024). The association between quality of therapeutic alliance and treatment outcomes in teletherapy: A systematic review and meta-analysis. Clinical Psychology Review, 110, 102430. https://doi.org/10.1016/j.cpr.2024.102430
American Psychological Association. (2025). Ethical guidance for AI in the professional practice of health service psychology. https://www.apa.org/topics/artificial-intelligence-machine-learning/ethical-guidance-ai-professional-practice
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/artificial-era-canonical-definition
Cross, S., Titov, N., Dear, B., Gleeson, J., & Alvarez-Jimenez, M. (2026). Psychological Therapy in the Age of Large Language Models: Framework for Therapist-Delivered and AI-Supported Functions. JMIR Mental Health, 13, e103712. https://doi.org/10.2196/103712
Diel, A., Torous, J., Cuijpers, P., Kleesiek, J., Nensa, F., Weber, N., Faust, F., Josan Lalgi, T., Mellis, F. S., Teufel, M., et al. (2026). A scoping review on the mental health harms of LLM-based chatbots. npj Digital Medicine, 9, 644. https://doi.org/10.1038/s41746-026-03054-x
Eubanks, C. F., Muran, J. C., & Safran, J. D. (2018). Alliance rupture repair: A meta-analysis. Psychotherapy, 55(4), 508–519. https://doi.org/10.1037/pst0000185
Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. https://doi.org/10.1037/pst0000172
Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI, 2(4). https://doi.org/10.1056/AIoa2400802
Hua, Y., Siddals, S., Ma, Z., Galatzer-Levy, I., Xia, W., Hau, C., Na, H., Flathers, M., Linardon, J., Ayubcha, C., & Torous, J. (2025). Charting the evolution of artificial intelligence mental health chatbots from rule-based systems to large language models: a systematic review. World Psychiatry, 24(3), 383–394. https://doi.org/10.1002/wps.21352
Kuang, J., Pope, A. L., & Zhang, Y. (2026). Psychotherapists’ Trust, Distrust, and Generative AI Practices in Psychotherapy: Qualitative Study. Journal of Medical Internet Research, 28, e88932. https://doi.org/10.2196/88932
Lee, H., Handler, R., Mungle, T., & Hernandez-Boussard, T. (2026). Building safer artificial intelligence mental health chatbots: a framework for transparency, evaluation, and shared accountability. Journal of the American Medical Informatics Association, 33(8), 1538–1553. https://doi.org/10.1093/jamia/ocag078
Olisaeloka, L., Richardson, C. G., Wang, A. Y., Munthali, R. J., & Vigo, D. V. (2026). Safety Mechanisms and Risk Mitigation in Generative AI Mental Health Chatbots: A Systematic Scoping Review. Healthcare, 14(10), 1395. https://doi.org/10.3390/healthcare14101395
Pichowicz, W., Kotas, M., & Piotrowski, P. (2025). Performance of mental health chatbot agents in detecting and managing suicidal ideation. Scientific Reports, 15, 31652. https://doi.org/10.1038/s41598-025-17242-4
Wang, Q., Zhong, J., Qi, K., Bao, H., Wang, Y., Guo, S., & Chen, A. (2026). The effects of artificial intelligence-based chatbots on mental health: A systematic review and three-level meta-analysis. Psychiatry Research, 366, 117459. https://doi.org/10.1016/j.psychres.2026.117459
World Health Organization. (2025). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. https://www.who.int/publications/i/item/9789240084759
World Health Organization. (2026). Towards responsible AI for mental health and well-being: experts chart a way forward. https://www.who.int/news/item/20-03-2026-towards-responsible-ai-for-mental-health-and-well-being--experts-chart-a-way-forward
Xiao, N., He, L., Chen, L., Liu, S., Xiang, Q., & Wang, P. (2026). Evidence on artificial intelligence-assisted clinical documentation and healthcare workers’ emotional wellbeing at work: a scoping review. Frontiers in Psychology, 17, 1840884. https://doi.org/10.3389/fpsyg.2026.1840884
Yang, H., Chang, F., Muroi, F., Liu, Z., Zhang, W., & Cai, J. (2026). Commercial AI-Based Mental Health Chatbots as Low-Intensity Adjuncts to Psychotherapy: Effectiveness, Adherence, and Safety – A Systematic Review and Meta-Analysis. Psychotherapy and Psychosomatics. https://doi.org/10.1159/000552072